Rule Extraction from Neural Networks－A Study for Financial Performance Prediction of Taiwan Listed Companies

• Main Authors: Xu-Sheng Jian, 簡旭生
• Other Authors: 左杰官
• Format: Others
• Language: zh-TW
• Published: 2006
• Online Access: http://ndltd.ncl.edu.tw/handle/53305601791709606813

碩士 === 國防管理學院 === 資源管理研究所 === 94 === This research project investigates the ability of neural networks for financial performance prediction and classification. We attempt to present a formal study on the complex phenomenon of financial performance to a sample of Taiwan listed companies for the period of 2000–2005. The predictor attributes include 6 macroeconomic variables and 12 financial statement variables and 5 ownership structure variables. The rate of return on equity is used as the to-be-predicted variable. In this article, we use three different periods of financial data to compare the prediction accuracy of the rate of return on equity. The experimental results show that one season’s financial data is more accuracy than others. In addition, we also use three different types of input variables to compare the prediction accuracy of the rate of return on equity. The experimental results show that one season’s “financial data + ownership structure” is more accuracy than “financial data + macroeconomic data” and “financial data + macroeconomic data + ownership structure”. Finally, we apply TREPAN algorithms to extract three binary decision trees from three propagation neural networks which include the prediction of return on equity and the prediction of earning per share and the classification of the rate of returns on stock price. We successfully present the prediction logic of neural networks and operation procedure of black box with decision trees, and find the important independent variables and very accurate knowledge in each binary decision tree.